A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
fuse fp32 for GLM-4.5 e_score_correction_bias (#22143)
Signed-off-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Yuxuan Zhang committed
Aug 4, 2025 at 00:04 UTC
d3c18c9cb0b6c42eab4ed7251adbf68dde4da39a
1 file changed
+2
-3
vllm/model_executor/models/glm4_moe.py
+2
-3
@@ -125,9 +125,8 @@ class Glm4MoE(nn.Module):
125
quant_config=None,
126
prefix=f"{prefix}.gate")
127
128
- # noaux_tc is not set in transformers new config now
129
- self.gate.e_score_correction_bias = (nn.Parameter(
130
- torch.empty(config.n_routed_experts)))
128
+ self.gate.e_score_correction_bias = nn.Parameter(
129
+ torch.empty(config.n_routed_experts, dtype=torch.float32))
130
131
# Load balancing settings.
132
vllm_config = get_current_vllm_config()